What the history reveals
New capabilities broaden the bundle. Performance gains deepen it. Connections and access determine how much of it people can actually employ.
Counts describe this curated release inventory, not market share, the size of a breakthrough, or a measured capability index.
Milestones by type
Each milestone has one primary classification. “Range + performance” is counted once.
Milestones by provider
All represented providers are shown here, including those without a timeline filter button.
How the mix changes over time
| Year | Range | Performance | Both | Access | Connection | Total |
|---|
2022 begins November 30; 2026 ends September 30. These are unequal observation windows, not annual growth rates.
The range-expansion narrative
The opening bundle was already substantial. ChatGPT’s launch made conversation a widely accessible doorway into writing, reasoning, and coding assistance. Image generation, speech processing, and emerging music generation also existed. November 2022 is the start of this story, not the invention of every capability in it.
During 2023, the assistant acquired ways to reach, calculate, and see. External tools connected it to information and services. Executable analysis turned an answer into calculations, charts, and files. Image understanding brought visual work into the conversation, while accessible video generation broadened what people could create.
During 2024, perception became interaction—and interaction became action. Memory could carry context forward. Live conversation could respond as a person spoke. Computer use could act through an interface. Research agents could pursue a question across sources. Live video and screen sharing brought changing visual context into the exchange.
From 2025 onward, the story increasingly becomes what these capabilities can accomplish together. Scheduled actions add a time dimension. Video with native audio combines media capabilities. Coding environments, cloud agents, and general agents connect reasoning to execution and deliverables. Many later milestones strengthen these workflows without adding another broad capability category.
The range additions represented in the graph
Ten additional categories join the seven-category starting bundle. These are categories used by this illustration, not a complete census of AI capabilities or claims of first-ever invention.
Does range actually plateau?
The rectangle’s width does plateau late in this timeline. Under its current categories, the last new broad category is scheduled actions in January 2025. Later range-related releases may add a subtype or combine categories already present, such as native audio in generated video.
That is partly a feature of the classification. “Coding,” “image generation,” and “computer use” are large buckets. A system that can sustain a much longer software task, preserve an identity across image edits, or reliably complete an entire workflow may unlock materially new work while remaining inside the same bucket.
It is also a feature of the illustration: width follows category additions, and height follows classified performance milestones. The visual therefore supports a useful conceptual story, but it cannot establish that real-world capability growth has slowed or become purely vertical.
Better performance can expand practical range.
A task that was theoretically possible but too slow, fragile, or expensive can become usable after a performance improvement. The category stays the same; the set of work people can realistically delegate expands.
Other patterns worth noticing
These observations use the Full inventory, regardless of the summary view selected above.
Performance dominates the later record.
Across the inventory, 70 of 104 milestones are performance-only improvements. In the 2026 portion, that is 21 of 26. This is a shift in the selected release mix—not a measurement of the amount of capability gained.
Connections deserve their own story.
Seventeen milestones are connections or integrations, and five are access changes. Together they account for about 21% of the inventory. Packaging existing capabilities into a workflow can materially change what a person can do, even without a new model capability.
The record is concentrated.
OpenAI, Anthropic, and Google account for 88 of 104 entries—about 85%. They are prominent in this inventory, but this also reveals its selection bias. Release counts should not be treated as a provider ranking or proof that other developers contributed less.
The overhang can grow inside a fixed category.
Someone may still use AI for email drafting while its coding, visual, research, and execution capabilities improve. Even when category count stops rising, the gap between relevant capability and employed capability can keep widening.
Research snapshot: September 30, 2026. This is a curated history, not an exhaustive audit. Preview access, plan and regional restrictions, and release limitations remain documented on individual timeline milestones. The source links support the releases; the classifications and synthesis are editorial interpretations.